Stratified sampling exists to stop chance from under-representing groups that matter.
It uses knowledge you already have about the population, such as how many customers live in each country.
Each stratum is sampled separately, so its share of the sample is controlled rather than left to luck.
Proportional allocation gives each stratum the same share of the sample as it has of the population.
For example, a population of 90% UK and 10% Irish customers gives 900 and 100 in a sample of 1,000.
Disproportionate allocation oversamples small strata, such as taking 500 from each country, to get reliable estimates for both.
Overall estimates then need weights, so each Irish customer counts for less than each UK customer in the combined result.
The weights undo the oversampling, so each stratum counts in proportion to its true share of the population.
Strata work best when members within a stratum are similar and the strata differ from each other.
Stratifying on a variable unrelated to the question adds work without making estimates more precise.
Every member must belong to exactly 1 stratum, so the strata must cover the population without overlap.
The stratum of each member must be known before sampling, which needs a richer list than simple random sampling.
In practice surveys often stratify by region, age group or customer type.
Stratifying by a rare outcome, such as cancelled orders, makes sure that outcome appears in the sample at all.
The usual mistake is oversampling a small group and then forgetting to weight it back when reporting the overall figure.